Zuckerberg manifesto pushes open-source approach on AI as Meta releases latest model

Meta Platforms announced the release of a new AI model, Muse Glimmer, for developers to use freely. CEO Mark Zuckerberg expressed concerns about the risks of concentrated control over advanced AI in the hands of a select few entities, including companies, institutions, and governments. The company provided a permissive open-source license for Muse Glimmer, which can run on personal computers. Zuckerberg also unveiled Muse Spark 1.2, a more powerful AI model, accessible to developers.

Meta has a history of open-sourcing AI models, but has faced challenges keeping pace with rivals like OpenAI, Anthropic, and Google. In an online essay, Zuckerberg outlined his vision for AI development and advocated for open-source access as the preferred approach for fair and broadly beneficial progress. He warned that if AI superintelligence were controlled by a few individuals or entities, it would likely lead to less favorable outcomes for others.

Zuckerberg emphasized the need to distribute advanced AI broadly to empower billions of people. He also addressed potential competition from rivals, stating that if leading labs focused on building AI for corporations, governments, or other institutions, the balance of power would favor larger organizations over individuals. The U.S. and allied nations must accelerate the development of energy capacity and infrastructure to remain competitive against rivals like China.

To address concerns about Chinese AI companies using distillation to extract technical features from U.S. models, Zuckerberg supported the principle of distillation, which involves training a less capable model on the outputs of a stronger one. He argued that learning from observations is a fundamental aspect of the world and the U.S. cannot lead if it restricts itself in this area.